The libc tree had no math.h. POSIX has 206 mandatory math functions. After landing PRs #5 and #6 earlier in the session, I went looking for the biggest single-sprint expansion target on the scorecard, and 206 cpp:no entries in one category was hard to ignore. The result is a greenfield libm slice — bit-manipulation IEEE 754 only, no transcendentals — that flipped 96 entries on each kernel and kicked both past 62% strict POSIX 1003.1-2024 in one commit.

Why bit-manip first

The libc-wide CFLAGS include -mgeneral-regs-only, which forbids floating-point in any libc translation unit — every existing libc file is FP-free for a reason. To get math working at all, the build needed a per-file escape hatch. Rather than open the whole flag wall at once, I added a compile_c_fp() helper to build.sh that strips -mgeneral-regs-only and adds -msse2 -mfpmath=sse -mlong-double-64 for math.c only. Same archive (libjefec.a), no link-time changes for callers, and the 80-bit x87 long double is sidestepped by aliasing it to double (the *l variants exist as real symbols for POSIX scoring but use double precision).

What 96 entries buys you

Thirty-three functions, three variants each (double/float/long double), plus twelve classification and comparison macros. The functions split cleanly into bit-cast categories: sign manipulation (fabs, copysign), min/max with NaN-quiet vs NaN-propagating semantics (fmin/fmax + the POSIX 2024 fminimum/fmaximum + fdim), rounding (floor/ceil/trunc/round + l/llround for integer return + rint/lrint/llrint for current-mode + nearbyint), IEEE 754 decomposition (frexp/ldexp/modf/scalbn/scalbln/logb/ilogb), the NaN constructor, nextafter/nexttoward, and fmod via musl-style mantissa-space repeated subtraction. The classification macros (signbit/fpclassify/isnan/isinf/isfinite/isnormal) dispatch to __signbit/__fpclassify helpers; the comparisons (isgreater/isless/...) compile to __builtin_* for guaranteed NaN-quiet semantics.

End-to-end on real hardware

The mathtest harness exercises every category once with the obvious traps: NaN propagation, signed zeros, halfway rounding (banker's rule for rint/nearbyint), subnormals via frexp's 2^54 normalization trick, fmod with mixed signs. Built it as a 243 KB static binary, embedded it into JefeRust's EMBEDDED_PROGRAMS array (psftp wedges silently on files over 700 bytes, so VHD-style deploy isn't an option for ad-hoc testing), rebuilt the kernel, and ran exec /programs/mathtest over SSH. The reply: mathtest PASS (all subtests OK). Every line in math.c that the harness covers behaves correctly on a real Hyper-V Gen1 VM under JefeRust's task scheduler.

Same userspace, both kernels

Because both JefeOS C++ and JefeRust run identical userspace/libc/ binaries against the same INT 0x80 syscall ABI, every libc-only sprint flips both scorecards in lockstep. C++ JefeOS went from 55.63% to 63.18% strict (+7.55pp); JefeRust from 55.07% to 62.63% (+7.56pp). The 0.55pp gap between them is unchanged — header-only entries C++ scored generously plus a couple of partial-to-yes promotions JefeRust hasn't claimed yet. The data.json edits were 96 surgical line replacements per kernel — no whole-file reformatting, so the diff stays readable and merges cleanly with whatever's in flight elsewhere.

What's Next

• LIBM-PHASE-2 (transcendentals) — sin/cos/tan/asin/acos/atan/atan2 + sinh family + exp/exp2/expm1/log/log2/log10/log1p + pow/sqrt/cbrt/hypot, all × {d,f,l}. Probably ~80 more entries with vendored musl algorithms (range reduction + minimax polynomials). Would push both kernels past 70% strict in a single session.
• fenv — small surface (12 entries: feclearexcept/fegetenv/fegetround/...), but unlocks proper FE_TONEAREST/FE_OVERFLOW reporting that PHASE-1 stubbed.
• PR #5's open follow-ups: parallel-N=5 SSH regression, X25519 ephemeral hardcoded, TCP CONNECTIONS BTreeMap unbounded.
• Real fixes for the three remaining smoke failures (sockettest2 IPv6 ntop, timer-posix case 3 timing, transient stringtest).